Meta cloud plan becomes the flashpoint for a wider chip selloff
U.S. storage chip names have come under heavy pressure since late June. According to a report cited by Yicai, SanDisk, Micron Technology, Seagate Technology and Western Digital have each fallen more than 20% over the past several weeks. The immediate spark was a Bloomberg report that Meta plans to launch a cloud infrastructure business and sell excess AI compute capacity to external customers.
The market read that report as a warning sign on two fronts: possible compute oversupply and the risk that AI capital spending is nearing a peak. Storage chips, a core part of AI infrastructure, were pulled into the selloff as money moved out of hardware-linked names. The article says the drop is not only about short-term sentiment. It has also reopened questions about the long-standing growth case for memory suppliers.
Two readings of Meta Compute
The bearish interpretation is straightforward. If a large platform is starting to sell unused compute, some investors see that as a sign internal demand is not absorbing capacity as expected. Bloomberg reported that Meta plans a cloud business called Meta Compute to market excess AI compute to outside companies, and that fed doubts about how durable demand is across the broader AI supply chain.
The opposing view is that the move says more about utilization than retreat. Mark Zuckerberg publicly rejected the idea of compute oversupply on July 10, saying he does not know anyone in the industry who believes they have excess compute, and adding that cloud services have commercial potential. The article notes that Meta has not cut its capital expenditure plan, and that its stock rose 9% after the news. On that reading, selling compute is a way to improve GPU cluster utilization and recover cash flow after large infrastructure spending, not a signal that expansion has stopped.
Training demand is softening, inference demand is holding up
The report points to a split market rather than a broad collapse. Data cited by TMTPost shows rental prices for training-focused compute, including Nvidia B200-related capacity, have pulled back in stages recently. By contrast, rental prices for AI inference compute used by government clients, enterprises and traditional industries have remained stable.
That distinction matters for memory suppliers. High Bandwidth Memory, or HBM, benefited directly from the race to train large models and had been in short supply. If training compute rental pricing keeps easing, marginal HBM demand growth may slow. General-purpose DRAM and enterprise SSD products tied more closely to inference deployments may hold up better, as inference use cases continue to spread across enterprise and traditional industry settings.
The article says another key variable is whether the technology gap between top AI models and their challengers keeps narrowing. If that gap closes quickly, the case for aggressive compute and memory spending weakens, and high growth assumptions for storage chips face a reset.
Long-term supply contracts are changing the memory business
Memory has historically been shaped by a familiar cycle: boom, capacity expansion, price collapse, contraction and recovery. Contracts were often short, and pricing moved like a commodity market. That structure is now being challenged as cloud providers and AI data centers seek supply security through longer deals.
According to the report, buyers are increasingly signing three- to five-year agreements with memory manufacturers that include price ranges, minimum purchase volumes and customer deposits. Micron has disclosed its first five-year strategic customer agreement. A Wallstreetcn report cited in the article says Samsung Electronics is in talks with Google and Microsoft on long-term supply deals that include more than $10 billion in prepayments. Reuters, also cited in the piece, reported that Tencent signed a three- to five-year supply deal worth more than 20 billion yuan with ChangXin Memory Technologies.
The article says this model could reduce the severity of price crashes and help manufacturers protect margins. It also cites institutions including Goldman Sachs, which view memory as moving away from a standardized commodity model toward a more customized business.
Buyers gain supply visibility but lose flexibility
Those long-term contracts come with trade-offs. In the old model, customers could add inventory through the spot market during downturns and benefit from lower prices. Under long-term agreements, they need to lock up large sums in deposits or prepayments, which raises pressure on cash flow.
Minimum purchase clauses add another constraint. Even if end demand softens, buyers may still need to take contracted volume or face financial penalties. The article notes that long-term contracts work well in rising price cycles because they secure lower costs, but in falling markets they can leave customers buying above spot prices and carrying paper losses. Smaller cloud companies and AI startups may find it harder to secure premium memory supply under this system, pushing access further toward large incumbents.
Investors are shifting from AI faith to earnings proof
The article places the memory selloff in a broader market reset. In June, the combined market value of the Magnificent Seven fell by about $3 trillion in a single month, while Microsoft dropped 21.64% over the same period. Wall Street is now asking tougher questions about returns on hundreds of billions of dollars in capital expenditure.
For cloud providers, that means AI spending has to be justified through revenue growth. The article uses Microsoft as an example, saying its heavy AI infrastructure investment needs support from Azure revenue expansion. If AI compute rental prices keep falling, cloud gross margins could come under pressure, which may affect the pace and willingness of future capital spending. Chip suppliers face their own risk: if training demand slows in stages, delivery cycles for high-end chips may stretch and order sizes may shrink, raising inventory pressure and pulling down earnings expectations for coming quarters.
Capital is rotating aggressively across the chain. Amundi, cited in the report, said money has been moving from cloud providers toward AI hardware and storage. Morgan Stanley, by contrast, has observed flows moving from chip stocks into AI cloud providers. The article reads that apparent contradiction as a sign the market is still searching for the next pocket of defensible growth.
Demand remains large, but the valuation framework has changed
The article also points to data that argues against a full stop in AI infrastructure buildout. According to reporting cited from The Paper, global monthly storage chip sales reached a record $74.6 billion. Micron’s capital spending plan for fiscal 2026 is more than $25 billion, nearly double from the previous year.
That does not suggest AI infrastructure is finished. It suggests investors are becoming more selective about the slope of growth. Across compute chips, cloud services, memory and applications, valuation is moving away from belief-led pricing and toward earnings-led pricing. A market with temporary oversupply in training compute and steady demand in inference is likely to apply more specific standards to each segment rather than reward the whole chain equally.